Project title: Video De-Occlusion
Host Institution: Technical University of Munich (TUM)
Host Supervisor: Prof. Laura Leal-Taixé
Co-host Institution: Technion
Co-host Supervisor: Prof. Michael Lindenbaum
Summary project: Occlusion is an important and open problem in computer vision. Most of modern, visual recognition algorithms still suffer from occlusion. Making the algorithms robust to occlusion is challenging, as occlusion is the result of the imaging process and emerges from the projection of a three dimensional world onto a two dimensional image.
Robust recognition therefore needs to identify occluded image regions and then to amodally complete across those gaps. However, this project will go a different way and instead of the single image the project will consider spatiotemporal, integrative mechanisms underlying sequences of images. The project aims especially at fragmented occlusion which happens in scenes with trees, and bushes as occluders (see the illustration).
It is well known in the cognitive sciences but less known in computer vision that the temporal dimension of vision introduces important cues for robust visual recognition. Concepts such as visual persistence and anorthoscopic perception rely on motion percepts which are not available in single images.
The project will study new algorithms for object detection on synthetic and real image sequences that mimic these concepts.